Exclusive-SpaceX targets in-house GPUs as it warns investors of chip supply, costs 

SpaceX targets in-house GPUs as it warns investors of chip supply, costs 

SpaceX's AI Ambitions and Chip Manufacturing Plans

By Echo Wang, Jeffrey Dastin and Max A. Cherney

NEW YORK, April 23 (Reuters) - SpaceX may be tackling one of the biggest challenges in the chip business: manufacturing the keys to powering artificial intelligence called graphics processing units, or GPUs.

Investor Warnings and Capital Expenditures

Ahead of SpaceX's $1.75 trillion IPO expected this summer, the company has warned prospective investors of its big spending plans to develop AI and other technologies.

It lists "manufacturing our own GPUs" among the "substantial capital expenditures" it is undertaking, according to excerpts of its S-1 registration reviewed by Reuters. Companies file this document to the U.S. Securities and Exchange Commission to disclose their risks and finances before going public.

SpaceX did not immediately respond to a request for comment, and the size of the expected expenditure could not be determined.

The Terafab Project and AI Chip Development

The ambition follows work by SpaceX, its xAI unit and Tesla to jointly develop the Terafab, an advanced AI chip manufacturing complex that CEO Elon Musk is planning in Austin, Texas.

Though Musk has said the project would target chips for cars, humanoid robots and space-based data centers, many details - including the types of AI chips, such as GPUs, it would produce - have been unknown.

Comparison with Industry Approaches

There are a range of approaches for chips that power AI. For example, Nvidia largely makes GPUs, which are general purpose and good at performing a wide array of data crunching tasks. Alphabet's Google takes another approach with its tensor processing units (TPUs), which are tuned to perform specific functions, key to building AI models and running chatbots such as Anthropic's Claude.

It was unclear when SpaceX plans to manufacture its own chip and which companies - the Terafab developers or their partner Intel - would handle the fabrication technologies inside the plant.

Musk told Tesla analysts on Wednesday that by the time Terafab scales up, Intel's next-generation 14A manufacturing process "will be probably fairly mature or ready for prime ​time" and "seems like the right move."

It was also unclear if SpaceX, in its filing, used the term GPU as shorthand for AI processors generally.    

Chip Supply Concerns and Industry Challenges

Still, the previously unreported plans for GPU production come as SpaceX warned investors that it may not have enough chip supply to power its growth.

SUPPLY CONCERNS

"We do not have long-term contracts with many of our direct chip suppliers," SpaceX said in the S-1 registration. "We expect to continue sourcing a significant portion of our compute hardware from third-party suppliers, and there can be no assurance that we will be able to achieve our objectives with respect to TERAFAB within the expected timeframes, or at all."

Manufacturing Complexity and Industry Landscape

Manufacturing GPUs is not easy. Industry heavyweight Nvidia pioneered GPU design and, like much of the industry, outsources their manufacture to Taiwan's TSMC.

TSMC has spent billions of dollars and years developing its most advanced manufacturing processes, which for cutting-edge chips require exotic materials and executing more than a thousand steps with atomic precision. Its years of manufacturing billions of Apple’s iPhone chips have afforded it an enormous amount of the required hands-on experience to produce cutting-edge processors.

The chip industry, as it is organized, now splits steps such as fabricating, packaging and testing among several discrete companies. Musk has said the Terafab will handle each step of chip production, including the design as well.

(Reporting by Echo Wang in New York, Jeffrey Dastin and Max A. Cherney in San Francisco; Editing by Kenneth Li and Kim Coghill)